Clouds are significant barriers to the application of optical remote sensing images. Accurate cloud detection can help to remove contaminated pixels and improve image quality. Many cloud detection methods have been developed. However, traditional methods either rely heavily on thermal infrared bands or clear-sky images. When traditional cloud detection methods are used with Gaofen 4 (GF-4) imagery, it is very difficult to separate objects with similar spectra, such as ice, snow, and bright sand, from clouds. In this paper, we propose a new method, named Real-Time-Difference (RTD), to detect clouds using a pair of images obtained by the GF-4 satellite. The RTD method has four main steps: (1) data preprocessing, including transforming digital value (DN) to Top of Atmosphere (TOA) reflectance, and orthographic and geometric correction; (2) the computation of a series of cloud indexes for a single image to highlight clouds; (3) the calculation of the difference between a pair of real-time images in order to obtain moved clouds; and (4) confirming the clouds and background by analyzing their physical and dynamic features. The RTD method was validated in three sites located in the Hainan, Liaoning, and Xinjiang areas of China. The results were compared with those of a popular classifier, Support Vector Machine (SVM). The results showed that RTD outperformed SVM; for the Hainan, Liaoning, and Xinjiang areas, respectively, the overall accuracy of RTD reached 95.9%, 94.1%, and 93.9%, and its Kappa coefficient reached 0.92, 0.88, and 0.88. In the future, we expect RTD to be developed into an important means for the rapid detection of clouds that can be used on images from geostationary orbit satellites.

Super-resolution (SR) technology has shown great potential for improving the performance of the mapping and classification of multispectral satellite images. However, it is very challenging to solve ill-conditioned problems such as mapping for remote sensing images due to the presence of complicated ground features. In this paper, we address this problem by proposing a super-resolution reconstruction (SRR) mapping method called the mixed sparse representation non-convex high-order total variation (MSR-NCHOTV) method in order to accurately classify multispectral images and refine object classes. Firstly, MSR-NCHOTV is employed to reconstruct high-resolution images from low-resolution time-series images obtained from the Gaofen-4 (GF-4) geostationary orbit satellite. Secondly, a support vector machine (SVM) method was used to classify the results of SRR using the GF-4 geostationary orbit satellite images. Two sets of GF-4 satellite image data were used for experiments, and the MSR-NCHOTV SRR result obtained using these data was compared with the SRR results obtained using the bilinear interpolation (BI), projection onto convex sets (POCS), and iterative back projection (IBP) methods. The sharpness of the SRR results was evaluated using the gray-level variation between adjacent pixels, and the signal-to-noise ratio (SNR) of the SRR results was evaluated by using the measurement of high spatial resolution remote sensing images. For example, compared with the values obtained using the BI method, the average sharpness and SNR of the five bands obtained using the MSR-NCHOTV method were higher by 39.54% and 51.52%, respectively, and the overall accuracy (OA) and Kappa coefficient of the classification results obtained using the MSR-NCHOTV method were higher by 32.20% and 46.14%, respectively. These results showed that the MSR-NCHOTV method can effectively improve image clarity, enrich image texture details, enhance image quality, and improve image classification accuracy. Thus, the effectiveness and feasibility of using the proposed SRR method to improve the classification accuracy of remote sensing images was verified.